split up generated kernels.cpp #829

This patch splits up kernels.cpp and CUDAkernels.cpp into multiple translation units, one per kernel. This improves compilation times, see the PR #829. closes #516.
daphne-eu · Oct 5, 2024 · 529f0cb · 529f0cb
1 parent 8d614f0
commit 529f0cb
Showing 12 changed files with 131 additions and 94 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -32,6 +32,7 @@ set(CMAKE_BUILD_WITH_INSTALL_NAME_DIR ON)
 set(CMAKE_CXX_STANDARD 20 CACHE STRING "C++ standard to conform to")
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+set(CMAKE_OPTIMIZE_DEPENDENCIES 1)
 
 set(CMAKE_CXX_FLAGS_DEBUG="${CMAKE_CXX_FLAGS_DEBUG} -g")
 set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O3")

diff --git a/src/runtime/local/kernels/CMakeLists.txt b/src/runtime/local/kernels/CMakeLists.txt
@@ -18,20 +18,21 @@
 set(CMAKE_CXX_STANDARD 20 CACHE STRING "C++ standard to conform to")
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+find_package(Python3 REQUIRED COMPONENTS Interpreter)
 
 # The library of pre-compiled CUDA kernels
 if(USE_CUDA AND CMAKE_CUDA_COMPILER)
     set(CMAKE_CUDA_STANDARD 20)
     set(CMAKE_CUDA_STANDARD_REQUIRED ON)
+    execute_process(
+      COMMAND
+        ${Python3_EXECUTABLE} genKernelInst.py kernels.json
+        ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDAkernels
+        ${PROJECT_SOURCE_DIR}/lib/CUDAcatalog.json CUDA
+      WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/)
 
-    add_custom_command(
-            OUTPUT ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDAkernels.cpp ${PROJECT_SOURCE_DIR}/lib/CUDAcatalog.json
-            COMMAND python3 ARGS genKernelInst.py kernels.json
-                    ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDAkernels.cpp ${PROJECT_SOURCE_DIR}/lib/CUDAcatalog.json CUDA
-            MAIN_DEPENDENCY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/kernels.json
-            DEPENDS ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/genKernelInst.py
-            WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/
-    )
+  file(GLOB CUDA_CODEGEN_CPP_FILES CONFIGURE_DEPENDS
+      "${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDA*.cpp")
 
     set(PREFIX ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/CUDA)
     set(CUDAKernels_SRC
@@ -59,7 +60,7 @@ if(USE_CUDA AND CMAKE_CUDA_COMPILER)
             ${PREFIX}/Solve.cpp
             ${PREFIX}/Syrk.cu
             ${PREFIX}/Transpose.cpp
-            ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDAkernels.cpp
+            ${CUDA_CODEGEN_CPP_FILES}
             ${PROJECT_SOURCE_DIR}/src/runtime/local/vectorized/TasksCUDA.cpp
             ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/VectorizedPipeline.h
             ${PROJECT_SOURCE_DIR}/src/runtime/local/vectorized/WorkerGPU.h
@@ -77,24 +78,28 @@ if(USE_CUDA AND CMAKE_CUDA_COMPILER)
     set_target_properties(CUDAKernels PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${PROJECT_SOURCE_DIR}/lib)
 endif()
 
-add_custom_command(
-        OUTPUT ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/kernels.cpp ${PROJECT_SOURCE_DIR}/lib/catalog.json
-        COMMAND python3 ARGS genKernelInst.py kernels.json ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/kernels.cpp ${PROJECT_SOURCE_DIR}/lib/catalog.json CPP
-        MAIN_DEPENDENCY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/kernels.json
-        DEPENDS ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/genKernelInst.py
-        WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/
-)
+execute_process(
+  COMMAND
+    ${Python3_EXECUTABLE} genKernelInst.py kernels.json
+    ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/kernels
+    ${PROJECT_SOURCE_DIR}/lib/catalog.json CPP
+  WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/)
+
+file(GLOB CODEGEN_CPP_FILES CONFIGURE_DEPENDS
+     "${PROJECT_BINARY_DIR}/src/runtime/local/kernels/*.cpp")
+# message("CODEGEN_CPP_FILES: ${CODEGEN_CPP_FILES}")
 
 list(APPEND LIBS DataStructures IO BLAS::BLAS MLIRParser)
 
 set(PREFIX ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/)
 set(HEADERS_cpp_kernels
     ${PREFIX}/MatMul.h
 )
+
 set(SOURCES_cpp_kernels
         ${PREFIX}/MatMul.cpp
         ${PROJECT_SOURCE_DIR}/src/runtime/local/instrumentation/KernelInstrumentation.cpp
-        ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/kernels.cpp
+        ${CODEGEN_CPP_FILES}
         ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/CreateDaphneContext.cpp
         ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/Pooling.cpp
         ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/VectorizedPipeline.h
@@ -104,7 +109,9 @@ set(SOURCES_cpp_kernels
         ${PROJECT_SOURCE_DIR}/src/runtime/local/vectorized/WorkerCPU.h
         )
 # The library of pre-compiled kernels. Will be linked into the JIT-compiled user program.
-add_library(AllKernels SHARED ${SOURCES_cpp_kernels} ${HEADERS_cpp_kernels})
+add_library(KernelObjLib OBJECT ${SOURCES_cpp_kernels} ${HEADERS_cpp_kernels})
+set_target_properties(KernelObjLib PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${PROJECT_SOURCE_DIR}/lib)
+add_library(AllKernels SHARED $<TARGET_OBJECTS:KernelObjLib>)
 set_target_properties(AllKernels PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${PROJECT_SOURCE_DIR}/lib)
 
 if(USE_CUDA AND CMAKE_CUDA_COMPILER)
@@ -130,4 +137,5 @@ if(USE_HDFS)
         find_library(LIBHDFS3 NAMES libhdfs3.so HINTS ${PROJECT_BINARY_DIR}/installed/lib REQUIRED)
 endif()
 
+target_link_libraries(KernelObjLib PUBLIC ${LIBS} ${MPI_LIBRARIES} ${PAPI_LIB} ${HWLOC_LIB} ${LIBHDFS3})
 target_link_libraries(AllKernels PUBLIC ${LIBS} ${MPI_LIBRARIES} ${PAPI_LIB} ${HWLOC_LIB} ${LIBHDFS3})
diff --git a/src/runtime/local/kernels/Cartesian.h b/src/runtime/local/kernels/Cartesian.h
@@ -30,7 +30,7 @@ void cartesianSet(ValueTypeCode vtcType, Frame *&res, const Frame *arg, const in
     }
 }
 
-void cartesian(Frame *&res, const Frame *lhs, const Frame *rhs, DCTX(ctx)) {
+inline void cartesian(Frame *&res, const Frame *lhs, const Frame *rhs, DCTX(ctx)) {
     const size_t numRowRhs = rhs->getNumRows();
     const size_t numRowLhs = lhs->getNumRows();
     const size_t totalRows = numRowRhs * numRowLhs;

diff --git a/src/runtime/local/kernels/CreateFrame.h b/src/runtime/local/kernels/CreateFrame.h
@@ -30,8 +30,8 @@
 // Convenience function
 // ****************************************************************************
 
-void createFrame(Frame *&res, Structure **colMats, size_t numColMats, const char **labels, size_t numLabels,
-                 DCTX(ctx)) {
+inline void createFrame(Frame *&res, Structure **colMats, size_t numColMats, const char **labels, size_t numLabels,
+                        DCTX(ctx)) {
     std::vector<Structure *> colMatsVec;
     for (size_t c = 0; c < numColMats; c++)
         colMatsVec.push_back(colMats[c]);
@@ -46,4 +46,4 @@ void createFrame(Frame *&res, Structure **colMats, size_t numColMats, const char
         delete[] labelsStr;
 }
 
-#endif // SRC_RUNTIME_LOCAL_KERNELS_CREATEFRAME_H
+#endif // SRC_RUNTIME_LOCAL_KERNELS_CREATEFRAME_H
diff --git a/src/runtime/local/kernels/Group.h b/src/runtime/local/kernels/Group.h
@@ -117,7 +117,7 @@ template <typename VTRes, typename VTArg> struct ColumnGroupAgg {
     }
 };
 
-std::string myStringifyGroupEnum(mlir::daphne::GroupEnum val) {
+inline std::string myStringifyGroupEnum(mlir::daphne::GroupEnum val) {
     using mlir::daphne::GroupEnum;
     switch (val) {
     case GroupEnum::COUNT:

diff --git a/src/runtime/local/kernels/InnerJoin.h b/src/runtime/local/kernels/InnerJoin.h
@@ -73,7 +73,7 @@ bool innerJoinProbeIf(
 // Convenience function
 // ****************************************************************************
 
-void innerJoin(
+inline void innerJoin(
     // results
     Frame *&res,
     // input frames

diff --git a/src/runtime/local/kernels/Quantize.h b/src/runtime/local/kernels/Quantize.h
@@ -42,7 +42,7 @@ template <class DTRes, class DTArg> void quantize(DTRes *&res, const DTArg *arg,
     Quantize<DTRes, DTArg>::apply(res, arg, min, max, ctx);
 }
 
-void calc_quantization_params(float min, float max, float &scale, uint8_t &quantized_zero) {
+inline void calc_quantization_params(float min, float max, float &scale, uint8_t &quantized_zero) {
     // Make sure that 0 is included
     min = (min > 0) ? 0 : min;
     max = (max < 0) ? 0 : max;
@@ -63,7 +63,7 @@ void calc_quantization_params(float min, float max, float &scale, uint8_t &quant
     }
 }
 
-uint8_t quantize_value(float a, float scale, uint8_t quantized_zero) {
+inline uint8_t quantize_value(float a, float scale, uint8_t quantized_zero) {
     // Map
     float value = static_cast<float>(quantized_zero) + a / scale;
 

diff --git a/src/runtime/local/kernels/SetColLabels.h b/src/runtime/local/kernels/SetColLabels.h
@@ -30,7 +30,7 @@
 // Convenience function
 // ****************************************************************************
 
-void setColLabels(Frame *&res, const Frame *arg, const char **labels, size_t numLabels, DCTX(ctx)) {
+inline void setColLabels(Frame *&res, const Frame *arg, const char **labels, size_t numLabels, DCTX(ctx)) {
     const size_t numCols = arg->getNumCols();
     if (numLabels != numCols)
         throw std::runtime_error("the number of given labels does not match "
@@ -51,4 +51,4 @@ void setColLabels(Frame *&res, const Frame *arg, const char **labels, size_t num
     delete[] labelsStr;
 }
 
-#endif // SRC_RUNTIME_LOCAL_KERNELS_SETCOLLABELS_H
+#endif // SRC_RUNTIME_LOCAL_KERNELS_SETCOLLABELS_H
diff --git a/src/runtime/local/kernels/SetColLabelsPrefix.h b/src/runtime/local/kernels/SetColLabelsPrefix.h
@@ -30,7 +30,7 @@
 // Convenience function
 // ****************************************************************************
 
-void setColLabelsPrefix(Frame *&res, const Frame *arg, const char *prefix, DCTX(ctx)) {
+inline void setColLabelsPrefix(Frame *&res, const Frame *arg, const char *prefix, DCTX(ctx)) {
     const size_t numCols = arg->getNumCols();
     const std::string *oldLabels = arg->getLabels();
     std::string *newLabels = new std::string[numCols];
@@ -50,4 +50,4 @@ void setColLabelsPrefix(Frame *&res, const Frame *arg, const char *prefix, DCTX(
     delete[] newLabels;
 }
 
-#endif // SRC_RUNTIME_LOCAL_KERNELS_SETCOLLABELSPREFIX_H
+#endif // SRC_RUNTIME_LOCAL_KERNELS_SETCOLLABELSPREFIX_H
diff --git a/src/runtime/local/kernels/Stop.h b/src/runtime/local/kernels/Stop.h
@@ -18,4 +18,6 @@
 
 #include <runtime/local/context/DaphneContext.h>
 
-void stop(const char *message, DCTX(ctx)) { throw std::runtime_error(std::string("system stopped: ") + message); }
+inline void stop(const char *message, DCTX(ctx)) {
+    throw std::runtime_error(std::string("system stopped: ") + message);
+}
diff --git a/src/runtime/local/kernels/ThetaJoin.h b/src/runtime/local/kernels/ThetaJoin.h
@@ -365,8 +365,8 @@ template <> class ThetaJoin<Frame, Frame, Frame> {
     }
 };
 
-void thetaJoin(Frame *&res, const Frame *lhs, const Frame *rhs, const char **lhsOn, size_t numLhsOn, const char **rhsOn,
-               size_t numRhsOn, CompareOperation *cmp, size_t numCmp, DCTX(ctx)) {
+inline void thetaJoin(Frame *&res, const Frame *lhs, const Frame *rhs, const char **lhsOn, size_t numLhsOn,
+                      const char **rhsOn, size_t numRhsOn, CompareOperation *cmp, size_t numCmp, DCTX(ctx)) {
     ThetaJoin<Frame, Frame, Frame>::apply(res, lhs, rhs, lhsOn, numLhsOn, rhsOn, numRhsOn, cmp, numCmp);
 }
 #endif // SRC_RUNTIME_LOCAL_KERNELS_THETAJOIN_H